Use of Deep Learning to Estimate Five-Year Risk of Advanced AMD
By Lynda Seminara
Selected and Reviewed By: Neil M. Bressler, MD, and Deputy Editors
Journal Highlights
JAMA Ophthalmology, December 2018
Download PDF
Burlina et al. applied deep learning (DL) to fundus images from the Age-Related Eye Disease Study (AREDS) to automatically assess the severity of age-related macular degeneration (AMD) and estimate the five-year risk of progression to advanced-stage AMD. They found that the DL model’s performance was comparable to that of humans when the AREDS 4-step severity scale was used. Promising results were achieved with the AREDS 9-step severity scale (which normally requires highly trained graders) as well as for estimating five-year risk of progression.
For their study, the authors gathered information from the AREDS dataset to develop deep convolutional neural networks that were trained to provide detailed automated AMD grading. Algorithm performance was compared with results from a human grader and against a criterion standard (gradings from a fundus photograph reading center). Three methods for estimating five-year risk were developed: hard, soft, and regressed. Main outcomes were weighted κ scores and mean unsigned errors for estimating five-year probability of progression to advanced AMD. The study included 67,401 color fundus images from a total of 4,613 patients.
Analysis showed a weighted κ score of 0.77 for the 4-step scale and 0.74 for the 9-step scale. The overall mean estimation error for 5-year risk ranged from 3.5% to 5.3%. The error was smaller for lower-risk classes. Of the three methods, hard prediction performed best for all classes except those in which the soft prediction outperformed all and in which the regressed prediction outperformed all.
The authors noted the large imbalances among some of the severity classes: For instance, for the 9-step scale, 24,411 images were classified as step 1, and 1,160 images were classified as step 3. Nonetheless, they said, DL has the potential to assist physicians with detailed risk assessment and evaluation of disease progression during treatment. (Also see related commentary by Harpal S. Sandhu, MD, FRCSC, Ayman El-Baz, PhD, and Johanna M. Seddon, MD, ScM, in the same issue.)
The original article can be found here.